Electronic Thesis and Dissertation Repository


Doctor of Philosophy




Dr. Robert Sica


A new first-principle Optimal Estimation Method (OEM) to retrieve ozone number density profiles in both the troposphere and stratosphere using Differential Absorption Lidar (DIAL) measurements obtained at the Observatoire de Haute Provence (OHP) in France is described. The method is robust and applicable to any DIAL ozone lidar. The ozone retrievals are compared to ozonesonde measurements, and these comparisons show the profiles match within the measurement uncertainties. The OEM retrieval also successfully catches much of the structure seen by the ozonesondes. The OEM retrievals are compared with the traditional analysis, and for most heights the difference between the two methods is small. One main advantage of the OEM is that all available measurements from multiple channels as well as lidars are used in the retrieval, eliminating the need to merge or perform corrections on the raw measurement. Thus, the tropospheric and stratospheric lidar measurements can be used together to generate an ozone profile which extends from 2.5\,km to about 42\,km. The upper troposphere and the lower stratosphere (UTLS) coincides with the measurements overlapping region. In the UTLS, even small changes in the distribution of the greenhouse gases can result in large changes in the atmospheric radiative forcing. The OEM can significantly improve the our understanding of the UTLS by providing an ozone density profile with a well-defined statistical and systematic uncertainty budget in this region. A new state-of-the-art machine learning technique was developed to automatically classify raw (level 0) lidar measurements to remove bad scans, and to distinguish between clear sky measurements and measurements with traces of either clouds or aerosols. We have examined different supervised learning methods and found the random forest classifier, the support vector machine (SVM), and the gradient boosting trees could successfully classify our lidar data with more than 90\% accuracy score with the random forest classifier recommended because of its greater computational speed.